The developers of an online backgammon platform analyzed 170,716 moves from 6,383 matches played by 938 users between March 30 and October 7, 2026. The dataset excludes developer accounts and test data, focusing on games played primarily against the engine.
The engine evaluates positions by calculating the expected outcome in points, ranging from -3 to +3. This metric accounts for three types of victory: oin (1 point), mars (2 points), and cox (3 points). The engine uses a neural network to assess board states and determine the optimal move.
The engine identifies the best move in 43% of cases. The analysis measures human performance using a performance rating (PR), where a PR of 10 corresponds to an average loss of 0.02 points per move. The engine's accuracy is verified through rollouts, which simulate games to reduce evaluation noise.
The rules for the engine are based on the Russian Backgammon Federation standards, including specific constraints like the prohibition of blocking all 15 opponent checkers and the requirement to play the maximum possible moves. The engine must account for these rules at every intermediate step of a move.
